Background and purpose: Infections are the leading cause of morbidity and mortality in patients with chronic lymphocytic leukemia (CLL) and occur during and after treatment. When deciding on the type of CLL treatment, the risk of infections is typically assessed based only on age and comorbidities; therefore, there is a need to develop a predictive model that incorporates information from multiple data sources. However, training an effective machine learning model requires a large sample size. Patient/material and methods: In this study, we developed a machine learning approach using domain adaptation (DA) to predict the risk of severe infection during treatment in patients with CLL. We implemented a DA strategy using lymphoma patient data and compared it with a domain-specific (DS) strategy across multiple models. Results: The DA strategy outperformed the DS strategy across all models, with an odds ratio of 4.43 for infection risk between high-risk and low-risk groups, compared with an odds ratio of 3.69 for the best DS model and 2.27 for the CLL-IPI alone. Explainability analysis identified predictive features for both the DA and DS models, including medication data and biochemistry tests. Specifically, C-reactive protein levels and non-therapeutic drugs were common features identified by both DA and DS models, while the DA models relied more heavily on alimentary tract drugs, solvents and diluting agents, and antibacterial medications. Interpretation: These findings highlight the value of integrating data from different diseases (lymphoma) to improve predictions in a target disease (CLL), and represent a step toward data-driven identification of CLL patients at high risk of infection during treatment.
Parviz et al. (Thu,) studied this question.